Machine Learning Application for Facies Distribution Using Combined K-Means and Random Forest Algorithm: Case Studies from Central Sumatra Basin, Indonesia

A deep understanding of rock facies is crucial for the reservoir characterization process. In this study, rock facies clusterization based on several variations of well log measurement data—GR, SP, LLD, LLS, density, and neutron, will be performed. Firstly, elbow methodology is employed to explore the possibility of the total class of facies for 7 different wells. This method results in 3 centroid “K” for the majority of well which is indicating the number of groups of lithology—shale, sandstone, and shaly-sand, that exist in each well. Following this result, the K-means algorithm is adopted to distinguish the location of every variety of lithology with respect to the depth of formation. This clustering process is executed in an individual well which is considered to have the deepest and complete range of data; hence, the model can cover further prediction for the other 6 wells. Lastly, the prediction model is performed using a random forest algorithm, coupled with a randomized search method with 3 cross-validations to tune the hyper-parameter of the algorithm. This process offers 150 fitting models of prediction with satisfying prediction results of 0.99 overall.